self-instruct

Definition

A recipe in which the model generates its own instruction–response pairs from a small pool of seed tasks, which are then filtered and used to fine-tune it. The appeal is cost: hand-writing instruction data is slow, and this turns a few hundred examples into tens of thousands. It is a data-generation method for supervised fine-tuning, not a preference-learning method like RLHF — in practice the two are stacked, not compared.

Explainers using this term

• Building a Dataset in Practice — Collect, Clean, BlendSelf-Instruct: Aligning Language Models with Self-Generated Instructions